README.md

metadata

license: cc-by-4.0
task_categories:
  - object-detection
language:
  - en
tags:
  - computer-vision
  - object-detection
  - yolo
  - virtual-reality
  - vr
  - accessibility
  - social-vr
pretty_name: DISCOVR - Virtual Reality UI Object Detection Dataset
size_categories:
  - 10K<n<100K
dataset_info:
  features:
    - name: image
      dtype: image
    - name: objects
      struct:
        - name: class_id
          list: int64
        - name: center_x
          list: float32
        - name: center_y
          list: float32
        - name: width
          list: float32
        - name: height
          list: float32
  splits:
    - name: train
      num_bytes: 536592914
      num_examples: 15207
    - name: test
      num_bytes: 29938152
      num_examples: 839
    - name: validation
      num_bytes: 59182849
      num_examples: 1645
  download_size: 613839753
  dataset_size: 625713915
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
      - split: validation
        path: data/validation-*

DIgitial Social Context Objects in VR (DISCOVR): A Social Virtual Reality Object Detection Dataset

Dataset Description

DISCOVR is an object detection dataset for identifying user interface elements and interactive objects in virtual reality (VR) and social VR environments. The dataset contains 17,691 annotated images across 30 object classes commonly found in 17 top social VR applications & VR demos.

This dataset is designed to support research in VR accessibility, automatic UI analysis, and assistive technologies for virtual environments.

The entire dataset is available to download at once at https://huggingface.co/datasets/UWMadAbility/DISCOVR/blob/main/dataset.zip

Pretrained YOLOv8 weights are available at https://huggingface.co/UWMadAbility/VRSight

If you use DISCOVR in your work, please cite our work VRSight for which it was developed:

@inproceedings{killough2025vrsight,
  title={VRSight: An AI-Driven Scene Description System to Improve Virtual Reality Accessibility for Blind People},
  author={Killough, Daniel and Feng, Justin and Ching, Zheng Xue and Wang, Daniel and Dyava, Rithvik and Tian, Yapeng and Zhao, Yuhang},
  booktitle={Proceedings of the 38th Annual ACM Symposium on User Interface Software and Technology},
  pages={1--17},
  year={2025}
}

---

Dataset Summary

• Total Images: 17,691
  • Training: 15,207 images
  • Validation: 1,645 images
  • Test: 839 images
• Classes: 30 object categories
• Format: YOLOv8
• License: CC BY 4.0

Object Classes

The dataset includes 30 classes of VR UI elements and interactive objects:

ID	Class Name	Description
0	avatar	User representations (human avatars)
1	avatar-nonhuman	Non-human avatar representations
2	button	Interactive buttons
3	campfire	Campfire objects (social gathering points)
4	chat box	Text chat interface elements
5	chat bubble	Speech/thought bubbles
6	controller	VR controller representations
7	dashboard	VR OS dashboard
8	guardian	Boundary/guardian system indicators (blue grid/plus signs)
9	hand	Hand representations
10	hud	Heads-up display elements
11	indicator-mute	Mute status indicators
12	interactable	Generic interactable objects
13	locomotion-target	Movement/teleportation targets
14	menu	Menu interfaces
15	out of bounds	Out-of-bounds warnings (red circle)
16	portal	Portal/doorway objects
17	progress bar	Progress indicators
18	seat-multiple	Multi-person seating
19	seat-single	Single-person seating
20	sign-graphic	Graphical signs
21	sign-text	Text-based signs
22	spawner	Object spawning points
23	table	Tables and surfaces
24	target	Target/aim points
25	ui-graphic	Graphical UI elements
26	ui-text	Text UI elements
27	watch	Watch/time displays
28	writing surface	Whiteboards/drawable surfaces
29	writing utensil	Drawing/writing tools

Dataset Structure

DISCOVR/
├── train/
│   ├── images/        # 15,207 training images (.jpg)
│   └── labels/        # YOLO format annotations (.txt)
├── validation/
│   ├── images/        # 1,645 validation images
│   └── labels/        # YOLO format annotations
├── test/
│   ├── images/        # 839 test images
│   └── labels/        # YOLO format annotations
└── data.yaml          # Dataset configuration file

Annotation Format

Annotations are in YOLO format with normalized coordinates:

<class_id> <center_x> <center_y> <width> <height>

All coordinates are normalized to [0, 1] range relative to image dimensions.

---

Usage

With Hugging Face Datasets

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("UWMadAbility/DISCOVR")

# Access individual splits
train_data = dataset['train']
val_data = dataset['validation']
test_data = dataset['test']

# Example: Get first training image and its annotations
sample = train_data[0]
image = sample['image']
objects = sample['objects']

print(f"Number of objects: {len(objects['class_id'])}")
print(f"Class IDs: {objects['class_id']}")
print(f"Bounding boxes: {list(zip(objects['center_x'], objects['center_y'], objects['width'], objects['height']))}")

With YOLOv8/Ultralytics

First, download the dataset and create a data.yaml file:

path: ./DISCOVR
train: train/images
val: validation/images
test: test/images

nc: 30
names:
  0: avatar
  1: avatar-nonhuman
  2: button
  3: campfire
  4: chat box
  5: chat bubble
  6: controller
  7: dashboard
  8: guardian
  9: hand
  10: hud
  11: indicator-mute
  12: interactable
  13: locomotion-target
  14: menu
  15: out of bounds
  16: portal
  17: progress bar
  18: seat-multiple
  19: seat-single
  20: sign-graphic
  21: sign-text
  22: spawner
  23: table
  24: target
  25: ui-graphic
  26: ui-text
  27: watch
  28: writing surface
  29: writing utensil

Then train a model:

from ultralytics import YOLO

# Load a pretrained model
model = YOLO('yolov8n.pt')

# Train the model
results = model.train(
    data='data.yaml',
    epochs=100,
    imgsz=640,
    batch=16
)

# Validate the model
metrics = model.val()

# Make predictions
results = model.predict('path/to/vr_image.jpg')

With Transformers (DETR, etc.)

from datasets import load_dataset
from transformers import AutoImageProcessor, AutoModelForObjectDetection

# Load dataset
dataset = load_dataset("UWMadAbility/DISCOVR")

# Load model and processor
processor = AutoImageProcessor.from_pretrained("facebook/detr-resnet-50")
model = AutoModelForObjectDetection.from_pretrained("facebook/detr-resnet-50")

# Process image
sample = dataset['train'][0]
inputs = processor(images=sample['image'], return_tensors="pt")

# Note: You'll need to convert YOLO format to COCO format for DETR
# YOLO: (center_x, center_y, width, height) normalized
# COCO: (x_min, y_min, width, height) in pixels

---

Applications

This dataset can be used for:

• VR Accessibility Research: Automatically detecting and describing UI elements for users with disabilities
• UI/UX Analysis: Analyzing VR interface design patterns
• Assistive Technologies: Building screen readers and navigation aids for VR
• Automatic Testing: Testing VR applications for UI consistency
• Content Moderation: Detecting inappropriate content in social VR spaces
• User Behavior Research: Understanding how users interact with VR interfaces

Citation

If you use this dataset in your research, please cite our publication using DISCOVR, called VRSight:

@inproceedings{killough2025vrsight,
  title={VRSight: An AI-Driven Scene Description System to Improve Virtual Reality Accessibility for Blind People},
  author={Killough, Daniel and Feng, Justin and Ching, Zheng Xue and Wang, Daniel and Dyava, Rithvik and Tian, Yapeng and Zhao, Yuhang},
  booktitle={Proceedings of the 38th Annual ACM Symposium on User Interface Software and Technology},
  pages={1--17},
  year={2025}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

You are free to:

• Share — copy and redistribute the material
• Adapt — remix, transform, and build upon the material

Under the following terms:

• Attribution — You must give appropriate credit

Contact

For questions, issues, or collaborations:

• Main Codebase: https://github.com/MadisonAbilityLab/VRSight
• This Repository: UWMadAbility/DISCOVR
• Organization: UW-Madison Ability Lab

Acknowledgments

This dataset was created by Daniel K., Justin, Daniel W., ZX, Ricky, Abhinav, and the MadAbility Lab at the University of Wisconsin-Madison to support research in VR accessibility and assistive technologies.